6
1

Exact Fractional Inference via Re-Parametrization & Interpolation between Tree-Re-Weighted- and Belief Propagation- Algorithms

Abstract

Inference efforts -- required to compute partition function, ZZ, of an Ising model over a graph of NN ``spins" -- are most likely exponential in NN. Efficient variational methods, such as Belief Propagation (BP) and Tree Re-Weighted (TRW) algorithms, compute ZZ approximately minimizing respective (BP- or TRW-) free energy. We generalize the variational scheme building a λ\lambda-fractional-homotopy, Z(λ)Z^{(\lambda)}, where λ=0\lambda=0 and λ=1\lambda=1 correspond to TRW- and BP-approximations, respectively, and Z(λ)Z^{(\lambda)} decreases with λ\lambda monotonically. Moreover, this fractional scheme guarantees that in the attractive (ferromagnetic) case Z(TRW)Z(λ)Z(BP)Z^{(TRW)}\geq Z^{(\lambda)}\geq Z^{(BP)}, and there exists a unique (``exact") λ\lambda_* such that, Z=Z(λ)Z=Z^{(\lambda_*)}. Generalizing the re-parametrization approach of \citep{wainwright_tree-based_2002} and the loop series approach of \citep{chertkov_loop_2006}, we show how to express ZZ as a product, λ: Z=Z(λ)Z(λ)\forall \lambda:\ Z=Z^{(\lambda)}{\cal Z}^{(\lambda)}, where the multiplicative correction, Z(λ){\cal Z}^{(\lambda)}, is an expectation over a node-independent probability distribution built from node-wise fractional marginals. Our theoretical analysis is complemented by extensive experiments with models from Ising ensembles over planar and random graphs of medium- and large- sizes. The empirical study yields a number of interesting observations, such as (a) ability to estimate Z(λ){\cal Z}^{(\lambda)} with O(N4)O(N^4) fractional samples; (b) suppression of λ\lambda_* fluctuations with increase in NN for instances from a particular random Ising ensemble.

View on arXiv
Comments on this paper